Reply to ‘Mosaic loss of chromosome Y in leukocytes matters’

A revised Correspondence letter by Forsberg and Dumanski¹ (founders of CRAY Innovation) posits that methodological differences in our analysis of mosaic chromosome Y loss (mLOY) in a substantially larger sample set² could explain the lack of replication of a borderline association of mLOY with all-cause as well as overall cancer survival³. The authors offer an interpretive hypothesis of their association study that mLOY is mechanistically linked to disease risk and mortality through immune cell function^3–5. They suggest that (1) associations between mLOY in blood derived DNA and disease risk or mortality could be diluted by including buccal derived samples in the analysis, (2) inclusion of lower completion rate samples could have skewed the overall results and (3) an inappropriate threshold for log₂ R ratio (LRR) deviation was used for mLOY scoring. Herein, we address these concerns by performing a series of sensitivity analyses of our study. Our expanded analysis indicates limited tissue-specific differences in the causes or consequences of mLOY, no evidence that the inclusion of lower completion rate samples influenced results, and no differences in the association with mLOY and cancer risk or mortality under a range of LRR mLOY scoring thresholds.

To address the first issue of buccal cell samples attenuating evidence for an association with cancer risk and mortality, we performed analyses stratified by DNA source for both blood and buccal samples (Table 1). In total, 9,697 samples were in the blood-derived DNA analyses and 4,029 samples were in the buccal-derived DNA analyses. As in our published paper², we observed strong associations for both age and smoking on risk of developing mLOY; yet little difference in the patterns observed for the analyses of blood and buccal-derived DNA. This is the first report of an age and smoking effect on mLOY observed in buccal derived DNA, suggesting the impact of smoking and age on mLOY can be observed in at least two tissue sources. As in the published combined analysis, stratified analyses for cancer risk suggest limited evidence for an association in the prospectively collected samples, but a potential association in the samples collected after cancer diagnosis (Table 1). Due to ample potential for confounding by disease or treatment, great caution is warranted when interpreting the effect of mLOY on cancer risk in post diagnostic samples. As previously reported, we observed no effect of mLOY on all-cause mortality or cancer specific mortality in either the blood or buccal strata (Table 1). In addition, no evidence for heterogeneity between blood or buccal-derived DNA was observed in the cancer risk or mortality analyses. Similar results were observed for the association between the 14q32.13 locus (marker SNP rs2887399 near TCL1A) and mean chromosome Y LRR in blood-derived (β=0.0114, P-value=5.86×10⁻⁸) and in buccal-derived (β=0.0199, P-value=2.01×10⁻⁵) DNA. This trend of similar association results by DNA tissue type was also observed for loci from a recent larger genome-wide association study of mosaic Y loss (Supplementary Table 1)⁶.

Table 1.

Association of mLOY with age, smoking, cancer risk and mortality.

	OR/HR	95% CI	Association P-value	Heterogeneity P-value
Age at DNA Collection
Blood	1.16	(1.14–1.18)	<2.0×10−16	3.27×10−5
Buccal	1.10	(1.08–1.12)	<2.0×10−16
Current Smoking
Blood	3.07	(2.15–4.38)	6.1×10−10	0.045
Buccal	1.81	(1.25–2.64)	1.9×10−3
Cancer Risk (All)
Blood	1.26	(1.06–1.50)	0.010	0.358
Buccal	1.48	(1.10–1.99)	0.010
Cancer Risk (DNA before Dx)
Blood	1.17	(0.97–1.42)	0.100	0.704
Buccal	1.07	(0.70–1.62)	0.760
Cancer Risk (DNA at or after Dx)
Blood	1.43	(1.11–1.85)	0.006	0.442
Buccal	1.69	(1.20–2.37)	0.003
All-cause Mortality*
Blood	0.87	(0.73–1.03)	0.100	0.526
Buccal	0.99	(0.69–1.42)	0.960
Cancer Mortality*
Blood	0.83	(0.68–1.01)	0.060	0.258
Buccal	1.07	(0.72–1.58)	0.750

^*Indicates Hazard Ratios (HR) are reported in place of Odds Ratios (OR). Age and smoking models are adjusted for smoking/age, contributing study and ancestry. Cancer risk models are adjusted for age, smoking, contributing study and ancestry. Mortality models are for individuals from prospective cohort studies with case DNA collected ≥1 year before cancer diagnosis and adjusted for age, smoking, cancer diagnosis, contributing study and body mass index. OR=Odds Ratio; HR=Hazard Ratio; 95% CI=95 Percent Confidence Interval; Dx=Cancer Diagnosis.

In response to the second issue that our laboratory had insufficient experience in filtering genotyping data, we highlight our extensive experience with Illumina genotyping platforms. Our laboratory has genotyped hundreds of thousands of samples and we have accordingly built a robust quality control and quality assurance pipeline that strictly follows manufacturer recommendations and best practices leading to over 400 peer-review publications. In fact, it was an observation from our laboratory that first identified the presence of genetic mosaicism in a population sample of “healthy” individuals⁷. It is unclear why the Forsberg and Dumanski group had remarkably poor genotyping results using the Illumina Human 610 Quad array (~50% passing QC)⁸. All mLOY events reported by our group on the Illumina Human 610 Quad array had a qPCR validation rate of 100%. To further rule out any possibility of the inclusion of lower genotyping completion rate samples potentially skewing our overall association results, we restricted our analysis to samples with high completion rates (≥ 95%). This resulted in the exclusion of 174 (1.2%) samples, 15 of which were classified as having mLOY. Association results were remarkably similar after the exclusion (Table 2), with the main difference, slightly higher association P-values due to the reduced sample size. Forsberg and Dumanski also highlight a lower validation rate in our quantitative polymerase chain reaction (qPCR) validation (88%) compared to their validation rate of 100%. This is largely a result of our higher LRR threshold for calling mLOY. Of the mLOY samples in our validation run, 100% of samples with LRR ≤ −0.4 replicated as whole chromosome losses in the qPCR assay, suggesting equal levels of concordance as the Forsberg analysis³. Finally, recent international, large-scale genome-wide association study consortia have highlighted the robustness of genotyping arrays with respect to samples collected at different dates, DNA extracted using a variety of methods and across a variety of different genotyping platforms^9,10. The same Illumina intensity data used to accurately call genotypes in these studies is also used to robustly calculate LRR^7,11,12.

Table 2.

Comparison of association results from the published report with results using stricter genotyping QC (≥ 95% genotype completion rate) and a lower LRR calling threshold.

	Published Results				Stricter QC Results				LRR ≤ −0.4
	OR/HR	LCL	UCL	P-value	OR/HR	LCL	UCL	P-value	OR/HR	LCL	UCL	P-value
Cancer Risk	1.19	1.00	1.42	0.047	1.18	0.99	1.40	0.072	0.97	0.70	1.34	0.838
All-cause Mortality*	0.89	0.76	1.04	0.149	0.90	0.77	1.05	0.191	0.82	0.59	1.15	0.251
Cancer Mortality*	0.87	0.73	1.04	0.124	0.88	0.73	1.05	0.142	0.92	0.64	1.31	0.641

^*Indicates Hazard Ratios (HR) are reported in place of Odds Ratios (OR). All analyses use samples from prospective cohort studies with cancer cases collected ≥1 year before cancer diagnosis. Cancer risk model is adjusted for age, smoking, contributing study and ancestry. Mortality models are adjusted for age, smoking, cancer diagnosis, contributing study and body mass index. LCL=Lower Confidence Limit; UCL=Upper Confidence Limit.

For the third issue, Forsberg and Dumanski argue that our applied LRR scoring threshold is an inappropriate threshold for calling mLOY. While the Correspondence suggests the Forsberg manuscript³ was the first to use a continuous response variable in their analysis, only one analysis of all-cause mortality reported an association with mLRR-Y. The vast majority of reported results in the Forsberg analysis³ focus on an LRR threshold of ≤ −0.4 (~35% of cells) to call mLOY, whereas our analysis used a threshold of ≤ −0.15 (~23% of cells)². Forsberg describes their choice of threshold as “maximizing the phenotypic effect” and “certainly affect(ing) the results”. Such a method of selecting a LRR threshold could have the unanticipated effect of overtraining the analysis to the data, thus limiting the generalizability of the findings to external populations. Autosomal analyses by our group and others have robustly detected mosaic events affecting fewer than 10% of cells^11–13, indicating the ability to detect mosaicism at lower cellular fractions. To our knowledge, there is no convincing evidence that a specific threshold for mosaic proportion necessarily confers a reproducible and mechanistically proven effect. The haploid Y chromosome lacks B-allele frequency data (other than for the markers present in the relatively small pseudoautosomal and X-transposed regions homologous to the X chromosome), and thus, we agree that justifying the need to be more conservative when selecting an LRR threshold is an important ongoing scientific question, but exactly what LRR threshold or cellular proportion affected has not been established. Rather than using subjective LRR thresholds to call mLOY, we agree for the need to thoroughly explore different LRR thresholds as well as investigate LRR continuously as has been done in some recently published analyses assessing relationships with mLOY^6,14,15. To test for an effect of LRR threshold on our published cancer risk and mortality results, we conducted sensitivity analyses based on a LRR of ≤ −0.4. We found no difference in overall analytical conclusions when using this lower LRR threshold (Table 2).

In the Correspondence by Forsberg and Dumanski¹, an analysis of chromosome Y median LRR values is reported for a population of 121 men at 93 years of age, showing a statistically significant correlation between mLRR in blood and buccal derived DNA (Pearson r=0.36, P-value=5.9×10⁻⁵, data from Forsberg and Dumanski¹ Supplementary Table 1); albeit discordances are observed for select LRR mLOY calling thresholds (e.g., −0.4 and −0.15). This population of elderly men may exhibit survival biases and may not be generalizable to men that are decades younger, especially since the leukocyte progenitor cell populations become increasingly oligoclonal with age¹⁶. We likewise observe a stronger association of mLOY with age in blood-derived DNA (OR=1.16) versus in buccal-derived DNA (OR=1.10), suggesting further caution when interpreting the results of mLOY in elderly men. Forsberg and Dumanski demonstrate a correlation between mLOY in blood and buccal-derived DNA in elderly men. Assessing the strength of this correlation, especially in middle aged males, could be useful when designing future epidemiological studies of mLOY. It is notable that the analysis of nonagenarians could be an interesting subgroup, particularly since in the large autosomal surveys^11–13, the data suggests that mosaic events may tail off at older ages, but further ongoing studies will certainly be needed to address this. Furthermore, leukocyte contamination of buccal derived saliva samples is common and can account for approximately 30–60% of sampled cells in healthy adults and up to 80% of cells in cases of gingivitis¹⁷, further suggesting DNA from blood and buccal tissue types may show correlations in cells with mLOY. Therefore, in order to define the relevance of mLOY in leukocytes and/or specific leukocyte subpopulations with respect to other tissues, tissue samples with no blood cell contamination should be used.

In summary, we demonstrate the robustness of our previously published results regarding mLOY in blood and buccal-derived DNA. The importance of mLOY in men’s health will be born out in larger, future studies and not premature declarations¹⁸. We find limited differences in analyses stratified by blood and buccal tissues, as well as no difference in results based on the issues of QC stringency and LRR cutoff threshold raised by the Correspondence¹. So, there is not enough evidence to conclude that mLOY in leukocytes is more important than in other tissue sources based on current data. We end our re-analysis by pointing out that the periodic criticisms^1,4,5 amount to what the great English playwright, William Shakespeare penned, “Much Ado About Nothing”.